`nemo_rl.environments.nemotron_utils`#

Module Contents#

`_required_config_value`
`load_nemotron_video_model_config`
`_nemotron_video_target_resolution`	Return the SFT/vLLM-compatible `(width, height)` for a video frame.
`_resize_and_normalize_nemotron_video_frame`	Resize one frame with the same numeric order as stock vLLM.
`_flatten_nemotron_video_frame_messages`	Replace locally decoded frame items with ordered `<image>` markers.
`_render_nemotron_video_prompt`
`_expand_nemotron_video_placeholders`	Match vLLM’s timestamped one-wrapper-per-tubelet video replacement.
`process_nemotron_video_frames`	Port the source branch’s dynamic video-frame preprocessing contract.

nemo_rl.environments.nemotron_utils.NEMOTRON_VIDEO_PROCESSOR_NAMES#: ‘frozenset(…)’

nemo_rl.environments.nemotron_utils._required_config_value(config: Any, name: str) → Any#

nemo_rl.environments.nemotron_utils.load_nemotron_video_model_config(model_name: str) → Any#

nemo_rl.environments.nemotron_utils._nemotron_video_target_resolution( *, original_width: int, original_height: int, target_num_patches: int, patch_size: int, downsample_ratio: float, maintain_aspect_ratio: bool, ) → tuple[int, int]#: Return the SFT/vLLM-compatible (width, height) for a video frame.

nemo_rl.environments.nemotron_utils._resize_and_normalize_nemotron_video_frame( frame: PIL.Image.Image, *, target_height: int, target_width: int, norm_mean: torch.Tensor, norm_std: torch.Tensor, ) → torch.Tensor#: Resize one frame with the same numeric order as stock vLLM.

nemo_rl.environments.nemotron_utils._flatten_nemotron_video_frame_messages( messages: list[dict[str, Any]], ) → tuple[list[dict[str, Any]], list[PIL.Image.Image], list[int], float]#: Replace locally decoded frame items with ordered <image> markers.

nemo_rl.environments.nemotron_utils._render_nemotron_video_prompt( processor: Any, messages: list[dict[str, Any]], template_kwargs: dict[str, Any], ) → str#

nemo_rl.environments.nemotron_utils._expand_nemotron_video_placeholders( rendered_text: str, *, embeddings_per_tubelet: list[int], frame_indices: list[int], fps: float, temporal_patch_size: int, ) → str#: Match vLLM’s timestamped one-wrapper-per-tubelet video replacement.

nemo_rl.environments.nemotron_utils.process_nemotron_video_frames( processor: Any, messages: list[dict[str, Any]], *, template_kwargs: dict[str, Any], temporal_patch_size: int, target_num_patches: int, maintain_aspect_ratio: bool, ) → dict[str, torch.Tensor]#: Port the source branch’s dynamic video-frame preprocessing contract.